Analysis of Electronic Nose Data Using Logical Neurons

The object of this study was to evaluate the performance of networks of logical neurons in analysing data from the Warwick Electronic Nose. The results are compared to those previously obtained from a back-propagation network on the same data.The Warwick Electronic Nose consists of an array of twelve different tin oxide gas sensors, the response to an odour being determined from the conductivity changes of these sensors. Five different alcohol vapours were presented to the nose and the conductivities recorded in each case.The’ standard’ McCulloch and Pitts node used in most topologies — in particular back propagation — has proved successful in many applications. However, implementing nodes of this type in hardware is non-trivial due to the need to store, modify and multiply analogue variables. Logical or Boolean Neurons have their inputs and outputs limited to the set {0,1}. This allows easy implementation in hardware using RAM lookup tables. It is for this reason that Logical Neurons were selected as the basis for this study.In general, the logical neurons were less successful in classifying the alcohol data than a back-propagation technique. However, a 6 layer ω-state PLN performed almost as well with a 94% successrate compared to 100% for the MLP. Further work on logical neurons may lead to improvement by fully exploiting their capacity for generalisation.